Projects
Project pages for my research. Each page gives the idea, the evidence, and links to the paper and code.
From Noise to Diversity: Random Embedding Injection in LLM Reasoning
Random Soft Prompts: training-free random vectors appended to the input make reasoning branch early, settle later, and widen Pass@N. Also improves rejection-sampling SFT.
Project page →
ResidualQuant: KV Cache Quantization for Looped Transformers with 2-Bit Residuals
Looped Transformers share weights but not KV caches. ResidualQuant stores one INT4 anchor loop and 2-bit residuals for the rest, recovering BF16 accuracy with 80.7% less KV storage.
Project page (seal-kaist.github.io) ↗